Nitesh Shantha Kumar

Technical lead for a clearance and promotions platform with two ingestion paths — a synchronous REST API and an asynchronous Kafka consumer — plus a business-user application. I decomposed delivery into dependency-mapped worktracks so four engineers could build in parallel, and owned design review and integration across them.

Migrated clearance and end-of-lifecycle event publishing off a legacy Oracle and Airflow pipeline onto Kafka using a transactional outbox with confirmed-flush acknowledgement. It published to the existing topics with identical schemas, so downstream consumers needed zero configuration changes.

Cutover was gated on a production shadow-mode parity check behind split feature flags, with tiered rollback covering named failure scenarios. See Fig B.

Designed the replacement of ~28,000 blocking per-clearance product-service calls a day with a log-compacted Kafka topic consumed into a local read store, keeping the upstream service as a permanent fallback on a miss.

Before designing, I baselined production latency and error rate: 32ms average, p99 199ms, 0.05% errors across 197,579 calls. I wrote three architecture decision records, each naming the rejected alternative and the constraint behind it.

Chose asynchronous batch processing over synchronous downstream calls, then rate-limited to 2K clearances/min against a 10-minute SLA. That protects the upstream inventory, pricing and conflict services even with 5× headroom. Load tested to 2.5× the production limit.

The job loaded the entire queue each cycle against an upstream service that resolved ~7 records/sec, with per-request batching unavailable — a hard ceiling of about 350 rows per 50-second cycle. Try the numbers in Fig C.

I shipped bounded batching as containment, then per-call timeouts and commit-and-continue resilience as the structural fix, and published an RCA with five causal chains and 14 prioritised follow-ups.

Traced 5.6B sequential row reads with zero index scans to a COALESCE that blocked index use on a view’s join keys. Before proposing the fix, I proved the keys were never null and that row output was identical (EXPLAIN cost 82,460 → 13; rows scanned 7.5M → 3).

The fix removed the need for a proposed user-facing restriction.

Redesigned promotion redemption limit enforcement as a single atomic database operation that combines validation and counter increment, eliminating limit violations caused by race conditions. P99 redemption latency dropped from 1,000ms to 10ms, with support for 10× concurrency.

Built and published an installable plugin with 10 skills and 7 subagents that automates incident triage across Slack, New Relic, GitHub, Confluence, ServiceNow and Jira, backed by 36 structured root-cause documents. Another engineer has since extended it in a 1,557-line pull request.

Built the platform alerting programme after two incidents where monitoring looked healthy but could not fire: 11 category pages that pair every alert condition with concrete triage steps. 7 of the 9 proposed alerts are now live in production.